Generate Text

To interact with LLMs, you can use the generate_text function.


text = px.generate_text(prompt="Hello AI world!")

This function makes best effort to unify the API calls to different providers.

Features

Basic Usage
System Prompt
Messages (Multi-turn)
Structured Output (Pydantic)
Web Search
Custom Provider Model
Extensive Return

Check parameters section for all features.

Basic Usage


text = px.generate_text(prompt="Hello, world!")


Hello! How can I assist you today?

System Prompt

Use the system parameter to set the AI’s behavior and provide instructions.


text = px.generate_text(
  system="Only answer with the single digit number.",
  prompt="What is 2 + 2?",
)

Messages

Use messages for multi-turn conversations. Each message has a role (user or assistant) and content.


text = px.generate_text(
  system="No matter what, you must answer with 7.",
  messages=[
    {"role": "user", "content": "Hello AI Model!"},
    {"role": "assistant", "content": "7"},
    {"role": "user"

Structured Output (Pydantic)

Use response_format to get structured responses. You can pass a Pydantic model class and the response will be automatically parsed and validated.


from pydantic import BaseModel
 
class City(BaseModel):
    name: str
    country: str
    population: int
 
result = px.generate_text(
    prompt="What is the capital of France? Include population.",
    response_format=City
)
 
print(result.name)


Paris
France
2161000

You can also use more complex nested structures:


from pydantic import BaseModel
from typing import List
 
class Step(BaseModel):
    explanation: str
    output: str
 
class MathSolution(BaseModel):
    steps: List[Step]
    final_answer: str
 
result = px.generate_text(


Step 1: Multiply 25 by 4 to get 100
Step 2: Add 10 to 100 to get 110
Answer: 110

Web Search

Enable web_search to allow the model to search the web for up-to-date information. This is useful for questions about recent events or real-time data.


text = px.generate_text(
    prompt="What are the latest news about AI today?",
    web_search=True
)

Web search works best with models that support this feature natively (like some OpenAI and Anthropic models). ProxAI will automatically handle the capability based on the selected model.

Custom Provider Model

Set provider_model to a tuple of (provider, model) string values.


text = px.generate_text(
  prompt="Hello model! What is your name and which company do you belong to?",
  provider_model=('claude', 'opus')
)


My name is Claude and I was created by Anthropic. It's nice to meet you!

You can also use px.types.ProviderModel to specify the provider and model.


provider_model = px.models.get_model(
    provider='openai',
    model='o4-mini')
print(provider_model)
text = px.generate_text(
    prompt="Hello model! What is your name and which company do you belong to?",
    provider_model=provider_model
)


('openai', 'o4-mini')
I’m ChatGPT, a language model developed and maintained by OpenAI.

Check all available models on ProxAI Github Repo

Extensive Return

Set extensive_return to True to get more information about the API call. This returns a px.types.LoggingRecord object.


from pprint import pprint
 
response = px.generate_text(
  prompt="Hello model!",
  extensive_return=True
)
pprint(response)


LoggingRecord(query_record=QueryRecord(call_type=<CallType.GENERATE_TEXT: 'GENERATE_TEXT'>,
                                       model=(<Provider.OPENAI: 'openai'>,
                                              <OpenAIModel.GPT_3_5_TURBO: 'gpt-3.5-turbo'>),
                                       prompt='Hello model!',
                                       system=None,
                                       messages=None,
                                       max_tokens=100,
                                       temperature=None,
                                       stop=None,
                                       hash_value=None),
              response_record=QueryResponseRecord(response='Hello! How can I '
                                                           'help you today?',
                                                  error=None,
                                                  error_traceback=None,
                                                  start_utc_date=datetime.datetime(2025, 2, 13, 0, 38, 17, 573491, tzinfo=datetime.timezone.utc),
                                                  end_utc_date=datetime.datetime(2025, 2, 13, 0, 38, 17, 930739, tzinfo=datetime.timezone.utc),
                                                  local_time_offset_minute=-0.0,
                                                  response_time=datetime.timedelta(microseconds=357285),
                                                  estimated_cost=150),
              response_source=<ResponseSource.PROVIDER: 'PROVIDER'>,
              look_fail_reason=None)

You can also convert response to a dictionary.


from dataclasses import asdict
pprint(asdict(response))


{'look_fail_reason': None,
 'query_record': {'call_type': <CallType.GENERATE_TEXT: 'GENERATE_TEXT'>,
                  'hash_value': None,
                  'max_tokens': 1000,
                  'messages': None,
                  'prompt': 'Hello model!',
                  'provider_model': {'model': 'gpt-4',
                                     'provider': 'openai',
                                     'provider_model_identifier': 'gpt-4-0613'},
                  'stop': None,
                  'system': None,
                  'temperature': None},
 'response_record': {'end_utc_date': datetime.datetime(2025, 5, 3, 17, 16, 58, 827118, tzinfo=datetime.timezone.utc),
                     'error': None,
                     'error_traceback': None,
                     'estimated_cost': 30000,
                     'local_time_offset_minute': -0.0,
                     'response': 'Hello! How can I assist you today?',
                     'response_time': datetime.timedelta(microseconds=790699),
                     'start_utc_date': datetime.datetime(2025, 5, 3, 17, 16, 58, 36446, tzinfo=datetime.timezone.utc)},
 'response_source': <ResponseSource.PROVIDER: 'PROVIDER'>}

Parameters

px.generate_text() parameters:

Option	Type	Default Value	Description
`prompt`	`str`	`None`	The prompt to generate text from.
`system`	`str`	`None`	System prompt to the model that will be prioritized over user prompt.
`messages`	`List[Dict[str, str]]`	`None`	List of messages that represents the history of the conversation. • Each element needs to be a dictionary with `role` and `content` keys. • `role` must be either `user` or `assistant` in alternating order. • `prompt` parameter shouldn’t be set if `messages` parameter is set.